Hugging Face's logo

shrisha
/
CrisperWhisper2.0_large-ggml

Automatic Speech Recognition
GGUF
English
multilingual
whisper.cpp
whisper
ggml
speech-recognition
audio
verbatim
Model card Files
xet
 Community
CrisperWhisper2.0_large-ggml / implementation_plan.md
shrisha's picture
shrisha
Update docs: corrected checksums, rebuild script, token-layout notes
e8ed05e verified
preview code
|
Raw
History Blame Contribute Delete
2.6 kB

Implementation Plan - Convert CrisperWhisper 2.0 Large to GGML/GGUF for SmartScribe

Convert the Hugging Face model nyralabs/CrisperWhisper2.0_large to GGML format and quantize it to q8_0 for use with SmartScribe's whisper.cpp engine.

User Review Required

Model Download Size & Processing: The nyralabs/CrisperWhisper2.0_large weights are ~3 GB. Downloading the model, converting to GGML (f16 format ~3 GB), and quantizing to Q8_0 will require approximately 6-8 GB of temporary local disk space.

Format Note: SmartScribe uses whisper.cpp for Whisper architecture models, which expects GGML format (quantized as q8_0 binary). We will produce ggml-crisperwhisper2.0-large-q8_0.bin and provide full conversion instructions and catalog integration.

Proposed Steps

1. Model Download & Preparation

  • Download model configuration, tokenizer, and safetensors weights from nyralabs/CrisperWhisper2.0_large using huggingface-cli or Python into a local workspace directory (models/CrisperWhisper2.0_large).

2. GGML Conversion

  • Run native/whisper-vulkan-addon/third_party/whisper.cpp/models/convert-h5-to-ggml.py using Python (referencing whisper/ for mel-filter assets).
  • Output intermediate float16 binary ggml-crisperwhisper2.0-large-f16.bin.

3. Quantization

  • Build the quantize utility in native/whisper-vulkan-addon/third_party/whisper.cpp/build if not already built.
  • Quantize the f16 binary to ggml-crisperwhisper2.0-large-q8_0.bin (q8_0).

4. Catalog Integration & Verification

  • Verify shasum 256 and binary integrity.
  • Document conversion log in release/whisper-crisperwhisper2.0-large-ggml/CONVERSION.md.
  • Add catalog entry for whisper-crisperwhisper2.0-large-q8 in src/shared/models-catalog.ts (if enabled in settings).

Proposed Changes

[NEW] release/whisper-crisperwhisper2.0-large-ggml/CONVERSION.md

  • Conversion log documentation detailing environment, download steps, conversion script invocation, and shasum checksums.

[MODIFY] models-catalog.ts

  • Add entry for CrisperWhisper 2.0 Large (whisper-crisperwhisper2.0-large-q8) to AVAILABLE_LOCAL_MODELS.

Verification Plan

Manual Verification

  1. Verify ggml-crisperwhisper2.0-large-q8_0.bin creation and non-zero byte size (~1.5 GB).
  2. Test loading and transcribing a sample audio file (jfk.wav) using test-whisper.js / node script with the converted binary model.